About

I am Yi Ge (Ellen), currently a Ph.D. student in Computer Science at the University of Maryland, College Park. I am a fervent researcher with a deep passion for AI agents, agentic world models, and multimodal large language models. During my graduate in Carnegie Mellon University (CMU), I used to work in Robotics Institute and Electrical and Computer Engineering Department.

Research Interests

My research interests center on Multimodal Large Language Models and Agentic AI, with a focus on building intelligent systems that can perceive, reason, and act across modalities. I am particularly interested in leveraging Reinforcement Learning and LLM-based Agents to enhance decision-making and interaction in multimodal settings. Recently, I mainly tackle challenges in

  • LLM/VLM Agents, Reinforcement Learning, and Reward Modeling.
  • Multimodal Reasoning and Evaluation for Large Vision-Language Models.
  • Synthetic Annotated Data Generation and Compositionality (e.g., AIGC, Layout Diffusion, Dataset Diffusion).
  • Domain Adaptation, Meta Learning and Transferability (e.g., UDA, partial UDA).

News

  • [10, 2026]NewANTE (Anticipate, Then Execute), a code world model with internal belief for coding agents, is now on arXiv. [Paper | Website | HuggingFace]
  • [5, 2026-9, 2026]NewMy first-author paper PatternBloom: Empowering Agentic RAG with Externalized RL-Distilled Graph Patterns is accepted in NeurIPS 2026.
  • [10, 2025-2, 2026]NewMy co-first-author paper Multi-Crit: Benchmarking Multimodal Judges on Pluralistic Criteria-Following is accepted in CVPR 2026 as Highlight.
  • [1, 2025-7, 2025]NewMy first-author paper DRUM: Learning Demonstration Retriever for Large MUlti-modal Models is accepted in ACL 2025 (srw).
  • [12, 2024-1, 2025]NewMy first-author paper FlexDataset: Crafting Annotated Dataset Generation for Diverse Applications is accepted in AAAI 2025 as Oral Presentation.
  • [12, 2024-1, 2025]NewMy first-author paper Reducing Divergence in Batch Normalization for Domain Adaptation is accepted in AAAI 2025 as Oral Presentation.
  • [11, 2024]Served as the reviewer for ICLR 2025.
  • [2024-2025]My co-first author paper Sharpness-Aware Minimization for Adversarial Domain Adaptation is under revision for TNNLS.
  • [04, 2023]Served as the reviewer for BMVC 2023.
  • [06, 2022]Our paper SphereVLAD++: Attention-based and Signal-enhanced Viewpoint Invariant Descriptor is accepted in RA-L 2022.
  • [04, 2022]Served as the reviewer for BMVC 2021.

Scroll for earlier news

Selected Publications

(* = equal contribution; My scholar name: Yi Ge, Ellen Yi-Ge)

[Preprints]

ANTE: Anticipate, Then Execute – Learning Internal World Models for Coding Agents from Verified Beliefs

Yi Ge, Kangrui Ruan, Jifan Chen, Manish Prajapat, Christian Bock, Xiaoyang Wang, Rein Houthooft, Aritra Sengupta, Varun Kumar

arXiv 2026 Preprint.

LLM AgentAgentic RLWorld ModelCoding

PaperWebsiteHuggingFaceVideoX post

An agentic RL framework (code world model) with internal belief.

[Peer-reviewed Conference & Journal & Workshop]

PatternBloom: Empowering Agentic RAG with Externalized RL-Distilled Graph Patterns

Yi Ge, Heng Huang

NeurIPS 2026 The 40th Annual Conference on Neural Information Processing Systems.

LLM AgentAgentic RLAgentic SearchMemory

PaperCode

An agentic RAG framework that distills RL-discovered reasoning patterns into externalized skills.

Multi-Crit: Benchmarking Multimodal Judges on Pluralistic Criteria-Following

Tianyi Xiong (*), Yi Ge (*), Ming Li, Zuolong Zhang, Pranav Kulkarni, Kaishen Wang, Qi He, Zeying Zhu, Chenxi Liu, Ruibo Chen, Tong Zheng, Yanshuo Chen, Xiyao Wang, Renrui Zhang, Wenhu Chen, Heng Huang

CVPR 2026Highlight IEEE/CVF Conference on Computer Vision and Pattern Recognition.

MultimodalEvaluation

PaperCodeWebsiteDataset

Evaluating multimodal judges on their capacity to follow pluralistic criteria and produce reliable criterion-level judgments.

FlexDataset: Crafting Annotated Dataset Generation for Diverse Applications

Yi Ge, Leo Shawn

AAAI 2025Oral Published. 39th AAAI Conference on Artificial Intelligence.

MultimodalAIGCDiffusionVisual Perception

PaperCodeVideoPoster

A framework for generating high-fidelity synthetic and annotated datasets using a composition-to-data (C2D) paradigm.

Reducing Divergence in Batch Normalization for Domain Adaptation

Yi Ge, Mingjing Wu, Zhenhao Chen

AAAI 2025Oral Published. 39th AAAI Conference on Artificial Intelligence.

Domain AdaptationVisual Perception

PaperCode

A normalization paradigm for unsupervised domain adaptation that mitigates estimation shifts accumulated in batch normalization layers.

DRUM: Learning Demonstration Retriever for Large MUlti-modal Models

Yi Ge, Jiechao Gao, Wei Han, Wei Zhu

ACL 2025 Published. The 63rd Annual Meeting of the Association for Computational Linguistics. SRW.

MultimodalIn-Context LearningAgentic Search

Paper

Demonstration retrieval tailored for large vision-language models (LVLMs).

Sharpness-Aware Minimization for Adversarial Domain Adaptation

Zhenhao Chen(*), Yi Ge(*), Wei Wang, Tianyi Liang

TNNLS 2024 Under Revision. IEEE Transactions on Neural Networks and Learning Systems.

Domain Adaptation

PaperCode

A framework for adversarial domain adaptation that integrates Interchangeable Batch Normalization (InterBN) and sharpness-aware optimization.

SphereVLAD++: Attention-based and Signal-enhanced Viewpoint Invariant Descriptor

Shiqi Zhao, Peng Yin, Yi Ge, Sebastian Scherer

RA-L 2022 Published. IEEE Robotics and Automation Letters.

3DRoboticsVisual Perception

PaperBibTexCode

An advanced method for 3D place recognition using LiDAR data.

Conditional Link Prediction of Category-Implicit Keypoint Detection

Yi Ge, Rui Fan, Zechun Liu, Zhiqiang Shen

WACV 2022 Published. 26th Winter Conference on Applications of Computer Vision.

Visual Perception

PaperBibTexTalk

A network for semantic keypoint detection and link prediction in multi-class instances.

Learning 3D Segmentation from Sparse Annotations via Hierarchical Descriptors

Yi Ge, Rui Fan, Zechun Liu, Zhiqiang Shen

ITCA 2023 Published. 2nd International Conference on Information Technology and Computer Application.

3DVisual Perception

PaperDemoMini Demo

A approach that can simultaneously learn segmentation from sparse annotations via reasoning global-regional structures and individual-vicinal properties.

Importance-Aware Semantic Segmentation in Self-Driving with Discrete Wasserstein Training

Xiaofeng Liu, Yuzhuo Han, Yi Ge, Song Bai, Tianxing Wang, Xu Han, Site Li, Jane You, Jun Lu

AAAI 2020Oral Published. 34th AAAI Conference on Artificial Intelligence.

Visual Perception

PaperBibTex

Incorporates the importance-aware inter-class correlation in a Wasserstein training framework by configuring its ground distance matrix.

Projects

[Other Projects]

[Project] Parallel Density-Based Spatial Clustering

Yi Ge

Project Website

Optimizes the DBSCAN algorithm for large datasets using various parallel computing strategies.

Tools: OpenMP, SIMD, CUDA, MPI, Cython.

[Project] High Dynamic Range Imaging Pipeline

Yi Ge

Project Website

HDR imaging pipeline integrating image merging, white balance correction, and tone mapping

Awards and Honors

  • Dean's Fellowship, UMD
  • Graduation with Distinction, CMU
  • Dean’s Honor List for Academic Excellence, CMU
  • Kaggle Competition Master Tier - Silver Medal (Top 5%)
  • World Mind Olympics - Top 3
  • National Undergraduate Academic Competition - Golden Price
  • International Mathematical Contest in Modeling - Meritorious Winner
  • SJTU merit-based Scholarship
  • SJTU Outstanding Volunteer Service Award
  • SJTU Social Practice Award